A large-scale enterprise autonomous driving model system based on data penetration and correlation

Abstract

This invention discloses an enterprise autonomous driving big data model system based on data penetration and correlation, belonging to the field of enterprise management technology. It includes: an enterprise autonomous driving big data model layer, used for unified processing, modeling, and analysis of multi-source data related to funds, providing basic capability support for upper-layer business intelligent agent applications; and a business intelligent agent application layer, located above the enterprise autonomous driving big data model layer, used to build intelligent application functions for specific business scenarios with the support of the enterprise autonomous driving big data model capabilities, realizing automated and closed-loop management of fund supervision. This constructs an enterprise autonomous driving big data model system of "data-driven + intelligent execution + risk closed loop," significantly improving the automation, intelligence, refinement, and security of enterprise operation management, and making up for the shortcomings of existing technologies in process automation, full-link tracking, cross-business collaboration, and proactive risk prevention and control.

Description

Technical Field

[0001] This invention relates to the field of enterprise management technology, specifically providing an enterprise autonomous driving large-scale model system based on data penetration and correlation. Background Technology

[0002] In the operation of enterprises / organizations, business operations involve multiple systems, roles, task chains, and cross-departmental collaboration, exhibiting characteristics of high complexity, dynamic change, and strong dependencies. Under traditional enterprise management systems, various business events typically rely on manual triggering and process-driven execution, resulting in delayed responses, broken collaboration, and inefficient resource scheduling, making it difficult to meet the requirements of modern enterprises for real-time, intelligent, and automated operations.

[0003] Enterprise operations management faces core problems of "fragmented triggers, delayed responses, and inefficient collaboration": On the one hand, various business trigger points (such as fund changes, contract signing, risk warnings, abnormal indicators, etc.) are scattered in different traditional systems, lacking a unified perception and centralized scheduling mechanism, forming "trigger silos"; on the other hand, traditional systems are mostly based on preset processes or manual operations, which cannot realize automatic driving of the entire chain of task linkage and resource response from a single point of event, making it difficult to build closed-loop execution capabilities.

[0004] From a business perspective, the integration of artificial intelligence and enterprise management systems is accelerating, gradually evolving from "decision assistance" to "intelligent execution." However, current mainstream enterprise software (such as Kingdee and Yonyou) still focuses on process standardization and information management, lacking task-driven end-to-end collaborative capabilities and failing to achieve intelligent linkage across systems and business domains. Furthermore, their architecture is typically modular, lacking a unified task orchestration and intelligent agent scheduling mechanism, making it difficult to support autonomous driving operation management in complex business scenarios.

[0005] From a technical perspective, existing related products or systems can be mainly divided into the following three categories:

[0006] First, process management and ERP systems (such as SAP, Kingdee, Yonyou, etc.) mainly realize the solidification of business processes and data recording, but their triggering mechanisms rely on manual or fixed rules, lack task-driven and dynamic response capabilities, and cannot achieve automatic execution of "single-point triggering and full-chain linkage".

[0007] Second, data analysis and BI tools (such as Wind and various data platforms) focus on information display and analysis decision-making. They usually require human intervention for judgment and operation, and cannot convert analysis results into automated execution actions, lacking the integrated capability of "analysis-decision-execution".

[0008] Third, rule-driven risk control and monitoring systems, while capable of identifying specific anomalies or risk events, are mostly based on static rule engines, lacking the ability to integrate multi-source data and perform dynamic reasoning, and are unable to achieve cross-system task linkage and closed-loop handling after a risk is triggered.

[0009] In summary, the existing technology system has not yet formed a method and system that can support enterprise intelligent operation management decision-making such as "single-point event triggering, multi-agent collaborative decision-making, automatic orchestration and execution of full-link tasks and closed-loop feedback of results", and cannot meet the needs of enterprises for real-time response, intelligent scheduling and global collaboration in complex business environments. Summary of the Invention

[0010] The purpose of this invention is to build a large-scale enterprise autonomous driving model system that supports "one-point triggering and full-chain response". Through unified event perception, intelligent agent collaborative decision-making and dynamic task chain orchestration, it realizes full-process automation and intelligent upgrade from event triggering to execution closed loop.

[0011] This invention provides a core underlying relational model centered around a large-scale enterprise autonomous driving model built around financial data. It comprehensively integrates multi-source data such as accounts, assets, contracts, transactions, vouchers, and projects, enabling visualization, traceability, and analysis of the entire financial chain. This allows companies to break down business barriers, standardize employee behavior, and improve the automation rate of enterprise financial management, automatic risk identification capabilities, and cross-departmental collaboration efficiency. It forms a large-scale enterprise autonomous driving model system with a single trigger and a full-chain response, addressing the shortcomings of traditional operation and management models in terms of global perception, dynamic response, and intelligent decision-making capabilities in the context of continuous business expansion, increasingly complex fund flows, and highly diversified account and asset structures.

[0012] To achieve the above objectives, this invention focuses on completing the following five core tasks:

[0013] 1. Establish a unified and scalable fund classification and standardized processing system to achieve unified fund management across business scenarios;

[0014] 2. Achieve intelligent correlation and penetration capabilities between capital flow and multi-dimensional business corpus, forming a unified semantic data modeling method;

[0015] 3. Build full-chain penetration and reverse tracing capabilities for funds, supporting multi-level penetration;

[0016] 4. Build multi-source data fusion capabilities to enable enterprises to achieve multi-agent data sharing, intelligent collaboration, and closed-loop processing capabilities;

[0017] 5. Establish a proactive monitoring and dynamic risk management mechanism to achieve early warning and in-process control of risks.

[0018] This invention constructs a large-scale enterprise autonomous driving model system with "data correlation and penetration + automatic task driving + risk prevention and control + business closed loop" as its core mechanism. This breaks down the data silos and process islands between traditional software systems, enabling the entire process of fund flow to be visualized, traceable, and analyzable. This supports the enterprise's refined operation management and risk control, effectively supports the enterprise's digital transformation strategy, improves internal operational efficiency and customer service capabilities, and provides basic support for the enterprise to expand new business growth models and intelligent application scenarios.

[0019] The technical solution proposed in this invention is: a large-scale enterprise autonomous driving model system based on data penetration and correlation, comprising:

[0020] The Enterprise Autonomous Driving Big Model Layer is used to uniformly process, model, and analyze multi-source financial data, providing basic capability support for upper-layer business intelligent agent applications.

[0021] The business intelligence agent application layer, located above the enterprise autonomous driving big model layer, is used to build intelligent application functions for specific business scenarios with the support of the enterprise autonomous driving big model capabilities, so as to realize the automated and closed-loop management of fund supervision.

[0022] The enterprise autonomous driving big model layer includes the following modules: data preprocessing submodule, corpus association submodule, fund classification submodule, data full-link tracking submodule, and enterprise data association penetration dashboard; the business intelligent agent application layer includes the following modules: intelligent guidance and interactive question answering module, abnormal business behavior control module, and task automatic driving and closed loop module.

[0023] Preferably, the data preprocessing submodule is used to interface with enterprise data from financial systems, capital systems, business systems and external regulatory systems, and to perform data cleaning, format standardization, field mapping and time series alignment processing to form structured data that can be used for subsequent model analysis.

[0024] Preferably, the corpus association submodule is used to associate fund flow data with a multi-dimensional corpus. The corpus includes at least business contracts, personnel information, fee rules, compliance systems, approval records, and other text or structured data related to fund transactions. Through the enterprise autonomous driving big data model, semantic understanding and relationship modeling of fund data and corpus data are performed to achieve the association and identification of the business entities, business scenarios, and compliance elements behind the funds.

[0025] Preferably, the fund classification submodule is used to automatically classify and process the enterprise's full business fund flow based on a preset general fund classification framework; the fund classification module supports configuring and expanding customized fund classification rules based on the enterprise's fund management and supervision needs on the basis of the general classification framework, so as to ensure classification consistency while adapting to personalized fund supervision requirements under different business scenarios.

[0026] Preferably, the data end-to-end tracking submodule constructs a three-level unified data model of "account - business data - classification," assigns a unique identifier to each funding node, and relies on a relational storage built with Elasticsearch to achieve forward and reverse link tracking: forward, it can parse from the account level to the transaction and asset, and reverse, it can trace back the corresponding funding source from the asset; the system achieves high-precision tracking of multi-level links through a depth-first traversal algorithm, while supporting custom penetration levels, and providing auditable basis annotations for each related node (such as contract number, transfer voucher, etc.); the final output includes a full-link funding flow map, an asset-fund ownership comparison table, and reverse tracing results, realizing transparent visualization and auditable management of complex funding chains.

[0027] Preferably, the enterprise data association and penetration dashboard is used to aggregate and visualize the analysis results output by the enterprise autonomous driving big model, forming a multi-dimensional fund monitoring view; the dashboard supports display according to fund category, business dimension, time dimension and risk dimension, and supports generating custom fund monitoring dashboards based on model analysis results; the enterprise autonomous driving big model layer operates collaboratively through the above modules to achieve in-depth understanding, structured expression and unified modeling of fund data, providing basic analytical capabilities for fund supervision and risk control.

[0028] Preferably, the intelligent guidance and interactive question-and-answer module is used to understand the financial data and related corpus based on the enterprise autonomous driving big model, to support users to query financial status, fund flow and related business information in natural language, and return structured or visualized query results.

[0029] Preferably, the abnormal business behavior control module is used to continuously monitor the current fund transaction behavior based on the learning results of the enterprise's autonomous driving big model on the characteristics of historical fund behavior, actively identify abnormal fund flows, illegal transactions or high-risk behaviors, and trigger corresponding early warning, analysis and disposal processes.

[0030] Preferably, the task automatic driving and closed-loop module is used to automatically generate corresponding processing tasks when specific financial events, abnormal behaviors or risk signals are identified, and distribute the tasks to the corresponding responsible entities according to preset rules or model inference results, track the task execution status until the task is completed, thereby forming a closed-loop processing flow for fund supervision.

[0031] The beneficial effects of this invention are:

[0032] This invention constructs a large-scale enterprise autonomous driving model system for intelligent operation management decision-making, breaking through the existing management model that relies on manual operation or static processes, and achieving an overall capability upgrade of "automatic task driving + data correlation and penetration + business closed loop + intelligent risk prevention and control". Compared with existing technologies, the main advantages include:

[0033] 1. Replace traditional OA and manual processes to achieve full-process automation.

[0034] By using task-driven intelligent task scheduling and multi-agent collaborative execution, this invention can automate business processes such as fund management, task distribution, and feedback processing, achieving a transformation from manual intervention to agent-driven operation, significantly reducing manual operation costs and improving process efficiency.

[0035] Compared with traditional OA systems, this invention not only supports process management, but also realizes the automatic triggering and closed-loop execution of business logic and cash flow.

[0036] 2. Data correlation and penetration capabilities and multi-dimensional business integration capabilities

[0037] By leveraging a unified semantic data model, multi-source heterogeneous associations of fund flows, contracts, projects, expenses, employee behavior, customer and compliance data are constructed to build a computable semantic fund flow path, enabling full-link penetration and reverse tracing from the source of funds to their final use.

[0038] Compared to traditional static reports or single-system tracking, this invention can achieve real-time and accurate fund correlation analysis in cross-system and cross-business domain scenarios.

[0039] 3. Business closed loop and dynamic risk management

[0040] This invention embeds anomaly identification and risk prevention capabilities into business intelligent agent applications, and achieves dynamic risk management of "pre-event warning + in-event control" through real-time monitoring, task-driven triggering and automatic handling.

[0041] Compared to traditional ex-post risk control or fixed rule-based risk control, this invention can proactively discover potential anomalies and risk patterns and automatically trigger handling processes, significantly improving the timeliness and coverage of risk response.

[0042] 4. Business scope adaptation and intelligent agent data sharing capabilities

[0043] Through scalable classification rules and semantic modeling mechanisms, this invention can flexibly adapt to the operational management needs of different enterprises and business lines, while providing processing results and analysis data to other intelligent agents for use, supporting multi-business collaboration and reuse.

[0044] Compared to existing technologies where systems are isolated and interfaces are limited, this invention achieves standardized and highly compatible cross-business data sharing, providing fundamental support for enterprises to build a multi-agent collaborative enterprise operation ecosystem.

[0045] 5. High explainability and regulatory capacity

[0046] The entire process of fund flow, task execution, and risk management is visualized, and can generate, trace, and interpretable fund path maps and event execution records.

[0047] Compared with traditional systems that only display summary results or single-stage reports, this invention not only provides results, but also explains the logic and behavior patterns of fund flow, meeting the requirements of enterprise internal control, auditing and supervision.

[0048] In summary, this invention constructs a large-scale enterprise autonomous driving model system based on "data-driven + intelligent execution + risk closed loop" through unified semantic modeling, task-driven intelligent scheduling, multi-agent collaboration, and dynamic risk management. This significantly improves the automation, intelligence, refinement, and safety of enterprise operation and management, and makes up for the shortcomings of existing technologies in process automation, full-link tracking, cross-business collaboration, and proactive risk prevention and control. Attached Figure Description

[0049] Figure 1 This is a block diagram of the system modules of the present invention.

[0050] Figure 2 This is a logic diagram illustrating the technical implementation of the data preprocessing module in the system of this invention.

[0051] Figure 3 This is a schematic diagram of the core entities of the data corpus of the present invention.

[0052] Figure 4 This is a diagram of the overall architecture of the corpus association of the present invention.

[0053] Figure 5 This is a diagram illustrating the architecture of the multi-source corpus association - project / contract category items of the present invention.

[0054] Figure 6 This is a diagram of the financial classification submodule architecture of the system of the present invention.

[0055] Figure 7 This is an example of the intelligent guidance and interaction submodule function of the system of the present invention. Detailed Implementation

[0056] The invention will be further described below with reference to the accompanying drawings.

[0057] This invention proposes an enterprise autonomous driving big model system for enterprise intelligent operation management decision-making. The system has a core architecture of "enterprise autonomous driving big model layer + business intelligent agent application layer" to achieve full-process, full-scenario, intelligent control over the internal and external operation management processes of enterprises.

[0058] The technical solution is divided into two main levels, each containing multiple functional modules. These modules work together through clear data and control flows to complete the processing, analysis, monitoring, and closed-loop management of multi-source data. At the same time, a data storage and support module is provided to support the entire system in terms of data storage, security, and technical support. The overall technical architecture is shown in Figure 1.

[0059] I. System Overall Technical Architecture

[0060] 1. Enterprise Autonomous Driving Large Model Layer

[0061] As the underlying core capability layer of this invention, it is used for unified processing, modeling, and analysis of multi-source data related to funds, providing basic capability support for upper-layer business intelligent agent applications. This layer includes at least the following functional modules:

[0062] Data preprocessing module: Used to connect with enterprise data from financial systems, capital systems, business systems and external regulatory systems, and to clean, standardize, map fields and align time series data to form structured data that can be used for subsequent model analysis.

[0063] The corpus association submodule is used to associate fund flow data with a multi-dimensional corpus. This corpus includes at least business contracts, personnel information, fee rules, compliance systems, approval records, and other textual or structured data related to fund transactions. Through the enterprise autonomous driving big data model, semantic understanding and relationship modeling of fund data and corpus data are achieved, enabling the identification of the underlying business entities, business scenarios, and compliance elements related to the funds.

[0064] The Funds Classification Submodule is used to automatically classify all business fund flows of an enterprise based on a preset general funds classification framework. This module supports configuring and extending customized funds classification rules according to the enterprise's fund management and supervision needs, thereby ensuring classification consistency while adapting to personalized fund supervision requirements in different business scenarios.

[0065] The data end-to-end tracking submodule constructs a unified three-level data model of "account—business data—classification," assigning each funding node a unique identifier. It leverages a relational storage system built on Elasticsearch to achieve forward and reverse link tracing: forward tracing starts from the account and proceeds level by level to the transaction and asset; reverse tracing starts from the asset and traces back to the corresponding funding source. The system uses a depth-first search algorithm to achieve high-precision tracking of multi-level links, while supporting customizable penetration levels and providing auditable annotations for each associated node (such as contract numbers, transfer vouchers, etc.). The final output includes a full-link funding flow graph, an asset-fund ownership comparison table, and reverse tracing results, enabling transparent visualization and auditable management of complex funding chains.

[0066] Enterprise Data Association and Penetration Dashboard: This dashboard aggregates and visualizes the analysis results output by the enterprise's large-scale autonomous driving model, forming a multi-dimensional financial monitoring view. The dashboard supports displaying data according to fund type, business dimension, time dimension, and risk dimension, and also supports generating custom financial monitoring dashboards based on model analysis results.

[0067] The enterprise autonomous driving big model layer operates collaboratively through the above modules to achieve in-depth understanding, structured expression and unified modeling of financial data, providing basic analytical capabilities for financial supervision and risk control.

[0068] 2. Business Intelligence Agent Application Layer

[0069] like Figure 1 As shown, the business intelligence agent application layer sits above the enterprise autonomous driving big model layer. It is used to build intelligent application functions for specific business scenarios, supported by the enterprise autonomous driving big model capabilities, to achieve automated and closed-loop management of funds supervision. Examples include investment analysis intelligence agents and expense control intelligence agents. Each intelligence agent application includes at least the following modules:

[0070] Intelligent guidance and interactive Q&A module: Based on the enterprise's autonomous driving big data model's ability to understand financial data and related corpora, it supports users to query financial status, fund flow and related business information in natural language, and returns structured or visualized query results.

[0071] Abnormal Business Behavior Control Module: Based on the learning results of the enterprise's autonomous driving big data model on historical fund behavior characteristics, it continuously monitors current fund transaction behavior, proactively identifies abnormal fund flows, illegal transactions or high-risk behaviors, and triggers corresponding early warning, analysis and handling processes.

[0072] The automatic task-driven and closed-loop module is used to automatically generate corresponding processing tasks when specific financial events, abnormal behaviors, or risk signals are identified. Based on preset rules or model inference results, the tasks are distributed to the corresponding responsible entities, and the task execution status is tracked until the task is completed, thereby forming a closed-loop processing flow for fund supervision.

[0073] Overall Operation Mechanism Description: In the overall operation of this invention, enterprise-related data first enters the enterprise autonomous driving big model layer. After data preprocessing, corpus association, and fund classification, structured data analysis results are formed. These analysis results are used to generate an enterprise intelligent operation management dashboard and to provide model / data capability support to the business intelligence agent application layer. Based on the above model capabilities, the business intelligence agent application layer performs data querying, anomaly identification, task triggering, and closed-loop processing operations, thereby achieving pre-event prevention, in-event control, and post-event traceability management of fund flows.

[0074] Through the above technical architecture, this invention realizes full-link penetration analysis and intelligent supervision of enterprise data, and supports flexible expansion of various enterprise operation and management application scenarios under a unified technical framework.

[0075] II. Enterprise Autonomous Driving Large Model Layer (Model Capability Layer)

[0076] As the core engine at the system's bottom layer, it forms the foundation of the entire enterprise's autonomous driving large-scale model system. It adopts a closed-loop architecture encompassing "input-processing-output-assurance" (assurance includes data quality monitoring, model iteration optimization, and prompt word optimization). Its core role is to provide standardized and highly reliable underlying technical capabilities to upper-layer intelligent agent application development modules. This module comprises four core sub-modules, ordered by data flow and functional support logic: data preprocessing sub-module, corpus association sub-module, fund classification sub-module, and fund penetration sub-module. These sub-modules collaborate to achieve three core objectives: multi-source data processing, intelligent fund analysis, and business adaptation support. The specific technical implementation logic will be elaborated upon below.

[0077] 1. Data Preprocessing Submodule

[0078] Technical implementation logic (such as) Figure 2 (as shown)

[0079] (1) Multi-source data access: Supports connection to multiple data sources such as banking systems, enterprise ERP, OA, CRM, and contract management systems, and provides multiple access methods such as API, direct database connection, and file import (Excel, CSV, PDF);

[0080] (2) Data cleaning and standardization: The "rule filtering + AI processing" mechanism is adopted to filter duplicate data and invalid data (such as test transactions with an amount of 0), and to standardize non-standard fields (such as unifying the abbreviation of the counterparty's name to the full name and unifying the date format to YYYY-MM-DD), and output structured data for subsequent core capability modules to call;

[0081] (3) Data quality monitoring: Real-time monitoring of data access quality (such as data integrity, accuracy, and timeliness), triggering early warnings for data sources with a data missing rate >5% and an error rate >3%, and notifying technical personnel to handle the issue.

[0082] Functional outputs: Standardized structured data and early warning notifications for data source anomalies.

[0083] Key technical points: The above sub-modules are implemented by using a scheduled task to call the SpringBatch batch processing framework to abstract the input, processing, and output layers.

[0084] 2. Corpus Association Submodule

[0085] Technical implementation logic (such as) Figure 3 , Figure 4 , Figure 5 (as shown)

[0086] (1) Corpus construction and management: Relying on data preprocessing steps, four types of structured corpora are integrated, and core related fields are defined (such as "Employee ID" in the employee behavior corpus being associated with "Operator ID" in the transaction details, and "Contract Number" in the contract agreement corpus being associated with "Business Contract Number" in the transaction details); a new corpus update mechanism is added to support automatic synchronization of data from the enterprise's internal systems (OA, CRM, contract management system) according to scheduled tasks (such as every day at midnight) to ensure the timeliness of the corpus;

[0087] (2) Intelligent matching algorithm: It adopts a three-level mechanism of "rule matching (high priority) + vector retrieval + large model confidence verification". The rule matching layer strengthens the deterministic association logic: First, it achieves accurate association based on bank transaction reconciliation code. For deterministic businesses such as expense reimbursement forms, it directly completes the binding of fund flow and corresponding business through the unique mapping relationship between reconciliation code and business documents; Second, it has a built-in maintenance of counterparty-business correspondence rule library, presets the association relationship between known counterparties (such as specific travel service providers, suppliers) and target businesses, and supports rule visualization configuration and batch import; For non-standardized scenarios without reconciliation code and non-preset counterparties (such as unfamiliar counterparties, model (For the business notes), firstly, extract the counterparty information, the party information, and the text features of the transaction notes from the transaction records. Then, use the Sentence-BERT model to generate multi-dimensional feature vectors through vector encoding. Based on the FAISS vector retrieval engine, quickly obtain the Top 10 candidate matching results from a multi-source corpus (retrieval efficiency ≤500ms / record). Next, input the candidate results and the original transaction text into a large model. The large model calculates the matching confidence of each candidate result and selects the data with high and medium confidence levels as the final association results. After binding with the bank transaction records, the data is stored in the database.

[0088] (3) Validation and optimization of association results: Through triple validation of “time consistency + amount correlation + business logic rationality” (such as the transaction time must be within the contract validity period and the transaction amount must match the contract amount within ±5%), a new association result feedback mechanism is added to support manual screening of low confidence and marking incorrect associations and triggering incremental training of the model to continuously improve the association accuracy.

[0089] Related dimensions:

[0090] (1) Employee business behavior database: related transaction details (operator, transaction amount) and employee performance data (performance indicators, assessment results); related to employee related expenses, such as travel expense reports and expense report details (reimbursement amount, reimbursement items, actual amount incurred).

[0091] (2) Decision-making process library: Approval process (applicant, approval status, approval attachments, approval opinion) corresponding to related transaction details and business data.

[0092] (3) Contract Agreement Library: Details of related transactions and key elements of contracts (contract number, partner, subject matter of transaction, maximum amount, performance period).

[0093] Customer database: Links employee and customer data such as earnings, contributions, and transaction details.

[0094] 3. Funds Classification Submodule

[0095] Technical implementation logic (such as) Figure 6As shown): the model classifies tags by using the associated corpus and the stream's own notes, counterparties, and account information.

[0096] (1) Customized classification framework: Provides a visual drag-and-drop configuration interface, supports users to customize fund classification by three dimensions: "business line - business type - transaction scenario" (such as "proprietary investment - bond investment - spot bond transaction"), and has built-in commonly used classification templates for financial and investment companies to reduce configuration costs;

[0097] (2) Automatic classification algorithm: By training the mapping relationship between business characteristics (such as counterparty type, purpose of funds, asset target, and transaction amount range) and fund classification, the model accuracy is ≥95%; incremental training is supported.

[0098] (3) Dynamic adaptation: When users add new business types, the classification framework is automatically triggered to adapt and the model association rules are updated synchronously, so that bank statements can be associated with the correct fund classification.

[0099] Functional outputs: Customized fund classification dictionary, full business transaction classification results, and classification accuracy statistics report.

[0100] Key technical points: Classification is achieved through a dual-track approach combining rules and models, along with prompt word engineering.

[0101] 4. Data end-to-end tracking submodule

[0102] Technical implementation logic:

[0103] (1) Data modeling: Construct a three-level data model of "account-business data-classification" and assign a unique identifier to each fund node (such as account ID, transaction serial number, unique business data code, and classification).

[0104] (2) Link tracing: Based on Elasticsearch to store the fund flow relationship, the forward full link tracing (account → transaction → asset, such as tracing from funds to specific bonds) and reverse source tracing query (asset → transaction → account, such as querying the corresponding fund source from bonds in reverse).

[0105] (3) Penetration accuracy control: Supports user-defined penetration levels (such as first-level penetration to the counterparty, second-level penetration to asset details), and marks the basis of each node association (such as transaction contract number, fund transfer voucher).

[0106] Functional outputs: a full-chain flow map of funds, an asset-fund ownership comparison table, and reverse tracing query results.

[0107] III. Application Layer of Business Intelligence Agent

[0108] Based on the standardized underlying capabilities provided by the enterprise autonomous driving big model module, specialized applications are developed for core enterprise business scenarios to achieve full-process support from data intelligent analysis to business closed-loop processing. The core includes three major functional sub-modules, and the specific implementation logic is as follows:

[0109] 1. Intelligent guidance and interactive Q&A submodule

[0110] Technical implementation logic (such as) Figure 7 (as shown)

[0111] (1) Based on the corpus association completed by the enterprise autonomous driving big model, APIs and API orchestration are constructed and injected into the model tools to expand the capability boundary of the model;

[0112] (2) Construction of interactive prompt word library: Based on high-frequency business scenarios (such as "analyze the performance of the bond team" and "query the source of funds of a certain asset"), pre-set guiding prompt words and support general language expression;

[0113] (3) Intent recognition: The user input text is parsed through the prompt word engineering to extract the core intent (such as the user input "Why has the recent investment income decreased?", the intent is recognized as "Analysis of abnormal reasons"), and the corresponding prompt words and functional modules are matched;

[0114] (4) Card-based display: The output information (such as analysis results and query data) is encapsulated into structured cards, including core indicators, related evidence, and visual charts (such as trend charts and percentage charts), and supports card drag-and-drop and combined viewing.

[0115] Functional outputs: Natural language guided interface, card-based information display results, and voice-to-text interaction support.

[0116] 2. Abnormal Behavior Control Submodule

[0117] Technical implementation logic:

[0118] (1) Construction of abnormal rule base: Preset high-frequency abnormal behavior rules (such as illegal bond holding on behalf of others, high buying and low selling of asset management products, trading beyond authority, and funds flowing to illegal accounts), and define rule triggering conditions (such as "the purchase price of the same asset is 20% higher than the market average price and the selling price is 10% lower than the market average price within a short period of time, triggering a high buying and low selling warning").

[0119] (2) Real-time monitoring: Streaming computing frameworks (such as Flink) compare the fund penetration results, corpus correlation data and abnormal rule base in real time, and generate abnormal warnings when rule matching is triggered;

[0120] (3) Automated handling process: Pre-set abnormal handling process (e.g., mild abnormality → remind the responsible person, moderate abnormality → suspend transaction + risk control review, severe abnormality → freeze account + compliance investigation), the system automatically triggers the corresponding process, synchronously notifies relevant departments (e.g., risk control, compliance, and funds departments), and records the entire handling process.

[0121] Functional outputs: abnormal behavior early warning notification, handling process tracking record, and abnormal control statistical report.

[0122] 3. Task-driven and closed-loop submodule

[0123] Technical implementation logic:

[0124] (1) Task generation: Based on the abnormal results identified by the abnormal behavior control submodule, a structured task sheet is automatically generated, which includes the task type (such as "performance optimization" and "compliance rectification"), responsible department (such as fixed income department and risk control department), task objective (such as "increase bond business revenue by 10% within 1 month"), completion deadline, and related basis (such as abnormal analysis report and relevant transaction data).

[0125] (2) Task assignment: Integrates with Yixin IM system API to automatically push task orders to the responsible department's workbench, and supports assignment according to role permissions (such as department head, specific executor).

[0126] (3) Progress tracking: Through the task status feedback mechanism (such as manual updates by the executor and automatic capture of execution data by the system), the task progress (not started, in progress, completed, feedback received) is tracked in real time, and timeout reminders are supported;

[0127] (4) Closed-loop confirmation: After the task is completed, the system automatically verifies the execution results (such as whether the performance data meets the standards and whether the compliance issues have been rectified), generates a closed-loop report, and synchronizes it to the management.

[0128] Functional outputs: structured task sheet, task progress tracking dashboard, and closed-loop completion report.

[0129] IV. Data Storage and Support Module

[0130] A hybrid storage architecture is adopted: Elasticsearch stores the fund flow relationship and the corpus association relationship; relational databases (such as MySQL) store structured business data (such as task orders and monitoring indicators); and distributed file systems (such as HDFS) store unstructured data (such as contract attachments and approval documents).

[0131] Data security mechanisms: Data security is ensured by employing technologies such as data encryption (encrypted transmission and encrypted storage), role-based access control (different departments can only see corresponding business data), and operation log auditing.

Claims

1. A large-scale enterprise autonomous driving model system based on data penetration and correlation, characterized in that, include: The Enterprise Autonomous Driving Big Model Layer is used to uniformly process, model, and analyze multi-source financial data, providing basic capability support for upper-layer business intelligent agent applications. The business intelligence agent application layer, located above the enterprise autonomous driving big model layer, is used to build intelligent application functions for specific business scenarios with the support of the enterprise autonomous driving big model capabilities, so as to realize the automated and closed-loop management of fund supervision. The enterprise autonomous driving big model layer includes the following modules: data preprocessing submodule, corpus association submodule, fund classification submodule, data full-link tracking submodule, and enterprise data association penetration dashboard; the business intelligent agent application layer includes the following modules: intelligent guidance and interactive question answering module, abnormal business behavior control module, and task automatic driving and closed loop module.

2. The system according to claim 1, characterized in that: The data preprocessing submodule is used to interface with enterprise data from financial systems, capital systems, business systems, and external regulatory systems. It cleans, standardizes, maps fields, and aligns time series data to form structured data that can be used for subsequent model analysis.

3. The system according to claim 1, characterized in that: The corpus association submodule is used to associate fund flow data with a multi-dimensional corpus. The corpus includes at least business contracts, personnel information, fee rules, compliance systems, approval records, and other text or structured data related to fund transactions. Through the enterprise autonomous driving big data model, semantic understanding and relationship modeling of fund data and corpus data are performed to achieve the association and identification of the business entities, business scenarios, and compliance elements behind the funds.

4. The system according to claim 1, characterized in that: The fund classification submodule is used to automatically classify and process the full amount of business fund flow of an enterprise based on a preset general fund classification framework. The fund classification module supports configuring and expanding customized fund classification rules based on the enterprise's fund management and supervision needs on the basis of the general classification framework, so as to adapt to the personalized fund supervision requirements under different business scenarios while ensuring classification consistency.

5. The system according to claim 1, characterized in that: The data end-to-end tracking submodule constructs a unified three-level data model of "account - business data - classification," assigning a unique identifier to each funding node. It leverages a relational storage system built on Elasticsearch to achieve forward and reverse link tracing: forward tracing starts from the account and proceeds level by level to the transaction and asset; reverse tracing starts from the asset and traces back to the corresponding funding source. The system uses a depth-first traversal algorithm to achieve high-precision tracking of multi-level links, while also supporting custom penetration levels and providing auditable annotations for each associated node (such as contract numbers, transfer vouchers, etc.). The final output includes a full-link funding flow graph, an asset-fund ownership comparison table, and reverse tracing results, enabling transparent visualization and auditable management of complex funding chains.

6. The system according to claim 1, characterized in that: The enterprise data association and penetration dashboard is used to aggregate and visualize the analysis results output by the enterprise autonomous driving big model, forming a multi-dimensional fund monitoring view. The dashboard supports display according to fund category, business dimension, time dimension, and risk dimension, and supports the generation of custom fund monitoring dashboards based on model analysis results. The enterprise autonomous driving big model layer operates collaboratively through the above modules to achieve in-depth understanding, structured expression, and unified modeling of fund data, providing basic analytical capabilities for fund supervision and risk control.

7. The system according to claim 1, characterized in that: The intelligent guidance and interactive question-and-answer module is used to understand the financial data and related corpus based on the enterprise autonomous driving big model. It supports users to query financial status, fund flow and related business information in natural language and returns structured or visualized query results.

8. The system according to claim 1, characterized in that: The abnormal business behavior control module is used to continuously monitor current fund transaction behavior based on the learning results of the enterprise's autonomous driving big model on historical fund behavior characteristics, proactively identify abnormal fund flows, illegal transactions or high-risk behaviors, and trigger corresponding early warning, analysis and handling processes.

9. The system according to claim 1, characterized in that: The automatic task-driven and closed-loop module is used to automatically generate corresponding processing tasks when specific financial events, abnormal behaviors, or risk signals are identified. Based on preset rules or model inference results, the tasks are distributed to the corresponding responsible entities, and the task execution status is tracked until the task is completed, thereby forming a closed-loop processing flow for fund supervision.

Images